Skip to main content

​
NexaCompute Quick Start Guide

​
Today’s Mission (2025-11-04)

Generate full dataset (10k-100k samples) and attempt training on 2×A100.

​
🚀 Quick Commands

​
1. Validate Environment

./scripts/validate_environment.sh

​
2. Run QC Batch (1k samples, ~10 min)

# With tmux (recommended)
./scripts/tmux_data_gen.sh 1 1000

# Direct execution
python scripts/run_batch_generation.py \
    --config batches/teacher_gen_v1.yaml \
    --input data/raw/scientific_corpus_325M.jsonl \
    --num-batches 1 \
    --batch-size 1000

​
3. Run Full Generation (100k samples, ~16-18 hrs)

# Start in tmux session
./scripts/tmux_data_gen.sh 10 10000

# Monitor progress
tmux attach-session -t data_gen

# Detach: Ctrl+B then D

​
4. Apply Filtering

python -m nexa_distill.sample_gate \
    --input data/processed/distillation/generated_combined.parquet \
    --output data/processed/distillation/filtered_samples.parquet \
    --rejections data/processed/distillation/rejections.parquet \
    --report data/processed/distillation/filter_report.md

# View report
cat data/processed/distillation/filter_report.md

​
5. Launch Training

# With tmux (recommended)
./scripts/tmux_training.sh

# Monitor
tmux attach-session -t training

​
6. Complete Pipeline (automated)

./scripts/run_full_pipeline.sh 10 10000

​
📋 Tmux Cheat Sheet

CommandDescription
tmux lsList sessions
tmux attach -t SESSIONAttach to session
Ctrl+B then DDetach from session
tmux kill-session -t SESSIONKill session
Ctrl+B then [Enter scroll mode
qExit scroll mode

​
📁 Key Files

​
Configurations

  • batches/teacher_gen_v1.yaml - Generation config
  • nexa_train/configs/baseline_qlora.yaml - Training config

​
Scripts

  • scripts/tmux_data_gen.sh - Data generation launcher
  • scripts/tmux_training.sh - Training launcher
  • scripts/run_full_pipeline.sh - Complete pipeline
  • scripts/validate_environment.sh - Environment check

​
Core Modules

  • nexa_eval/rubrics/judge_f.py - Factual accuracy judge
  • nexa_eval/rubrics/judge_r.py - Reasoning & safety judge
  • nexa_distill/sample_gate.py - Quality filtering
  • scripts/run_batch_generation.py - Generation orchestrator

​
Documentation

  • docs/TODAY_EXECUTION_PLAN.md - Detailed execution plan
  • QUICK_START.md - This file

​
🔍 Monitoring

​
Check Generation Progress

# Attach to tmux
tmux attach-session -t data_gen

# Or tail logs
tail -f logs/data_gen/run_*.log

# Check manifest
tail -f data/processed/distillation/generation_manifest.jsonl

​
Check Training Progress

# Attach to tmux
tmux attach-session -t training

# Or tail logs
tail -f logs/training/run_*.log

# TensorBoard
tensorboard --logdir logs/tensorboard --port 6006

​
Check GPU Usage

watch -n 1 nvidia-smi

​
⚠️ Troubleshooting

​
API Rate Limits

Edit batches/teacher_gen_v1.yaml: 
max_concurrent_requests: 20  # Reduce from 50

​
OOM During Training

Edit nexa_train/configs/baseline_qlora.yaml: 
per_device_train_batch_size: 2  # Reduce from 4
gradient_accumulation_steps: 8   # Increase from 4

​
Tmux Session Exists

tmux kill-session -t data_gen  # or training

​
💰 Cost Estimates

TaskCostDuration
QC Batch (1k)$0.50-0.75~10 min
Full Gen (100k)$50-75~16-18 hrs
Training (2×A100)$9~3 hrs
Total$60-85~19-21 hrs

​
✅ Success Criteria

​
Data Generation

  • ✅ 100k samples generated
  • ✅ Judge-F mean ≥ 75
  • ✅ Judge-R mean ≥ 75
  • ✅ Acceptance rate ≥ 70%

​
Training

  • ✅ Training starts without errors
  • ✅ Throughput ≥ 12k tokens/sec
  • ✅ Training loss decreases
  • ✅ Checkpoints saved

​
🎯 Today’s Workflow

​
📞 Need Help?

See detailed execution plan: 
cat docs/TODAY_EXECUTION_PLAN.md
Check module documentation: 
ls -la docs/Overview_of_Project/
Quick Start: ./scripts/validate_environment.sh && ./scripts/run_full_pipeline.sh